Discussion Week 8. TA: Kyle Dewey. Tuesday, November 15, 11

Transcription

1 Discussion Week 8 TA: Kyle Dewey

2 Overview Exams Interrupt priority Direct memory access (DMA) Different kinds of I/O calls Caching What I/O looks like

3 Exams

4 Interrupt Priority Process 1 makes an I/O request Process 2 makes an I/O request While setting up Process 2 s request, Process 1 s request finishes

5 Interrupt Priority Process 1 I/O Start I/O Done Time -> Process 2 I/O Start

6 Prioritizing While servicing one interrupt, another interrupt occurred Question: should we finish our ISR or allow the other ISR to run immediately?

7 Prioritizing Which ISRs preempt which is potentially complex Preemption within the kernel can be complex Similar issues as with process preemption

8 Performing I/O I/O takes forever, so while we do I/O, we schedule something else....so how do we do I/O if we re doing something else?

9 DMA Controller Answer: special hardware Dedicated processor just for handling I/O The processor directly manipulates memory and the I/O device, bypassing the CPU

10 DMA Messages Read X bytes starting at Y part of disk into Z part of memory Write X bytes to disk starting at part Y of disk from Z part of memory

11 DMA in Action Time 1 Time 2 CPU Write 3b M: 1; D: 0 DMA CPU Schedules Process DMA Controller Does I/O Memory Memory x f o o x f o o Disk Disk a b c d f b c d

12 DMA in Action Times 3-4 Time 5 CPU Runs Process DMA Controller Does I/O CPU I/O Done DMA Memory Memory x f o o Disk f o o d x f o o Disk f b c d

13 DMA Issues Question: How does this work in with virtual memory? Question: How does this work with swappable pages?

14 I/O Types

15 Blocking/Nonblocking Blocking: wait until I/O is complete until returning Nonblocking: on each call, return what has been done so far

16 Nonblocking Question: Why is this useful? Consider a program that gets the sum of integers from a 1 GB file

17 Synchronous/ Asynchronous Synchronous: wait for I/O to complete before returning Asynchronous: do not wait for I/O to complete before returning Ideally, there is a callback to signal when the I/O is complete

18 Asynchronous vs. Nonblocking DMA is asynchronous Asynchronous is still all or nothing, like a typical synchronous blocking I/O Nonblocking can get partial results Note: Wikipedia lies

19 I/O Calls Synchronous Asynchronous Blocking Common Depends... Nonblocking Nonsensical Possible

20 But I/O is still slow...

21 Caching Idea: Do as much I/O in memory as possible file = open( filename, O_RDWR ); read( file, buffer, 50 );... // more code that changes buffer lseek( file, 0, SEEK_SET ); write( file, buffer, 50 ); close( file )

22 Caching Massive performance gains possible Real systems use this all the time Can be quite complex

23 UNIX Semantics A write() is immediately available to anything that calls read() How does this work with caching?

24 Consistency Cache is by definition a copy of the actual resource Any access to that resource must behave as if caching is not being performed They will be out of sync, but they cannot look out of sync

25 Example Revisited How does the OS handle this? file = open( filename, O_RDWR ); read( file, buffer, 50 );... // more code that changes buffer lseek( file, 0, SEEK_SET ); write( file, buffer, 50 ); strcpy( buffer, Poisoned buffer ); close( file )

26 The face of I/O

27 Variation

28 Variation I/O is extremely broad...yet we try to make it fit into a consistent interface Major differences in OSes are often seen in how they treat I/O

29 Broad Classes Character based: One character at a time is manipulated Examples: keyboards, mice, network cards Block based: Blocks of characters are manipulated at a time Examples: Hard drives, flash drives

30 Broad Classes Some manage not to fit into even these Example: tape drives A seek can take a long time - impractical under essentially any circumstance Character devices have no notion of seeking However, data is stored in blocks...

31 Floppy Drives Often used the same interface as tape drives More naturally a block device

32 Responding to I/O Interrupts or polling can be used Some support both Which is best depends on usage scenario Polling makes little sense for hard drives - why? Interrupts are not usually used for mice - why?

33 I/O Space I/O devices often have a limited amount of space they can work with internally If the OS fails to respond before this space fills, data can be lost Example: keyboards

34 Buffering Buffer: a place in memory to put this data Not *quite* the same as a cache Not a copy Note that a cache can also be a buffer, but a buffer is by definition not a cache